Training on a dataset containing over a million image

[lucidrains]

Training on a dataset containing over a million image

[dancrew32]

Very interested in this as well. It seems like with the more images you use, the better the output is, but it is unclear when there are diminishing returns. E.g. Would 2M images really be better than 1M?

@franclin, as you train on this dataset, please keep us updated! It would be extremely interesting to see how quickly you can get to a high quality GAN output. If you can prototype GANs faster by using really small images, getting a faster iteration time, that would be an interesting way to see if investing the time in a high resolution output >=1024px is worth it.

[lucidrains]

yea, this will work with a million images